Image data processing method, image data recognition method, training method for image recognition model, image data processing apparatus, training apparatus for image recognition model, and image recognition apparatus

Abstract

Implementations of the present specification provide an image data processing method, an image data recognition method, a training method for an image recognition model, an image data processing apparatus, an image data recognition apparatus, and a training apparatus for an image recognition model. During image data processing, data anonymization processing is performed on image data based on frequency domain transform to obtain anonymized image data of the image data. The obtained anonymized image data includes a subgraph data set. Each subgraph data in the subgraph data set corresponds to a different frequency. Then, image blending processing is performed on the obtained anonymized image data based on data augmentation, to obtain augmented anonymized image data. In some implementations, graph size alignment processing is performed on each subgraph data in the augmented anonymized image data, so that a size of each subgraph data after the graph size alignment processing is same as a size of the image data in original form.

What is claimed is: 
     
         1 . A method, comprising:
 performing data anonymization processing on image data based on frequency domain transform to obtain anonymized image data of the image data, the anonymized image data including a first subgraph data set, and each subgraph data in the first subgraph data set corresponding to a different frequency; and   performing image blending processing on the subgraph data in the anonymized image data based on data augmentation, to obtain augmented anonymized image data.   
     
     
         2 . The method according to  claim 1 , wherein the data augmentation includes one or more of:
 mixup data augmentation; or   instahide data augmentation.   
     
     
         3 . The method according to  claim 2 , wherein the data augmentation includes mixup data augmentation, a number of images used for the image blending processing is k, and a maximum weight coefficient for blending image is W max ; and
 the performing image blending processing on the subgraph data in the anonymized image data based on data augmentation includes:
 performing k-1 times of disarrangement processing on the first subgraph data set of the anonymized image data to obtain k subgraph data sets including the first subgraph data set; 
 constructing an image hypermatrix with a size of m*k based on the k subgraph data sets, wherein a first column in the image hypermatrix corresponds to the first subgraph data set, and m is a number of pieces of subgraph data in the first subgraph data set; 
 randomly generating a weight coefficient for each image in the image hypermatrix; 
 performing row normalization on weight coefficients of images in the image hypermatrix, so that a sum of weight coefficients of each row of images is 1, and a weight coefficient of each image is not greater than W max ; and 
 performing weighted summation on a row of images of the image hypermatrix to obtain a blended image hypermatrix with a size of m*1, wherein images in the blended image hypermatrix are the augmented anonymized image data. 
   
     
     
         4 . The method according to  claim 3 , wherein W max  is 0.55, and k is 3. 
     
     
         5 . The method according to  claim 1 , further comprising:
 performing graph size alignment processing on subgraph data in the augmented anonymized image data.   
     
     
         6 . The method according to  claim 1 , wherein the performing data anonymization processing on the image data based on frequency domain transform includes:
 performing local frequency domain transform processing on the image data to obtain at least one feature map, wherein each feature map of the at least one feature map includes a plurality of elements and corresponds to a data block in the image data, and each element of the plurality of elements corresponds to a frequency in the frequency domain;   constructing, by using elements corresponding to frequencies in the at least one feature map, frequency component channel feature maps corresponding to the frequencies, respectively; and   obtaining at least one target frequency component channel feature map from the frequency component channel feature maps through screening, to obtain the anonymized image data of the image data, wherein a target frequency component channel feature map of the at least one target frequency component feature map is a channel feature for image recognition.   
     
     
         7 . The method according to  claim 6 , further comprising:
 after the obtaining the at least one target frequency component channel feature map from the frequency component channel feature maps through screening,
 performing a first shuffle processing on a target frequency component channel feature map to obtain a first shuffled feature map; and 
 performing normalization processing on the first shuffled feature map. 
   
     
     
         8 . The method according to  claim 6 , wherein the obtaining the at least one target frequency component channel feature map from the frequency component channel feature maps through screening includes:
 obtaining the at least one target frequency component channel feature map from the frequency component channel feature maps through screening based on one or more of channel importance or a screening rule.   
     
     
         9 . The method according to  claim 7 , further comprising:
 after the performing normalization processing on the first shuffled feature map,
 performing channel blending processing on the first shuffled feature map that has been normalization processed; 
 performing a second shuffle processing on the first shuffled feature map that has been channel blending processed, to obtain a second shuffled feature map; and 
 performing normalization processing on the second shuffled feature map to obtain the anonymized image data of the image data. 
   
     
     
         10 . The method according to  claim 9 , wherein
 the performing channel blending processing on the first shuffled feature map includes:   performing channel blending on two adjacent frequency component channel feature maps in the first shuffled feature map.   
     
     
         11 . The method according to  claim 1 , wherein the image data includes face image data. 
     
     
         12 . The method according to  claim 1 , further comprising:
 obtaining the image data as a training sample and labels of the image data;   obtaining a weight coefficient for each piece of subgraph data in the augmented anonymized image data;   performing label blending processing on labels corresponding to subgraph data in the anonymized image data by using weight coefficients of the subgraph data in the anonymized image data; and   training an image recognition model by using the subgraph data in the anonymized image data and the labels of the subgraph data that have been label blending processed.   
     
     
         13 . The according to  claim 12 , wherein in response to that subgraph data in the image blending processing comes from a same category, labels of the subgraph data in the image blending processing remain unchanged, and in response to that at least a part of subgraph data in the image blending processing comes from different categories, non-zero values in labels of the subgraph data are adjusted to k non-zero values, and each non-zero value corresponds to a weight coefficient of a piece of subgraph data. 
     
     
         14 . The method according to  claim 1 , further comprising:
 obtaining the image data as an image to be recognized;   providing the anonymized image data to an image recognition model to predict a feature vector representation of the image data; and   determining, based on the predicted feature vector representation and a feature vector representation of a target object, whether the image to be recognized includes the target object.   
     
     
         15 . The method according to  claim 14 , wherein the image data includes face image data, and the target object includes face image data with identity information. 
     
     
         16 . A computer system, comprising:
 one or more processors;   one or more storage devices coupled to the one or more processors, the one or more storage devices, individually or collective, having executable instructions stored thereon, which when executed by the one or more processors, enable the one or more processors to, individually or collectively, implement acts including:   performing data anonymization processing on image data based on frequency domain transform to obtain anonymized image data of the image data, the anonymized image data including a first subgraph data set, and each subgraph data in the first subgraph data set corresponding to a different frequency; and   performing image blending processing on the subgraph data in the anonymized image data based on data augmentation, to obtain augmented anonymized image data.   
     
     
         17 . The computer system according to  claim 16 , wherein the data augmentation includes mixup data augmentation, a number of images used for the image blending processing is k, and a maximum weight coefficient for blending image is W max ; and
 the performing image blending processing on the subgraph data in the anonymized image data based on data augmentation includes:
 performing k-1 times of disarrangement processing on the first subgraph data set of the anonymized image data to obtain k subgraph data sets including the first subgraph data set; 
 constructing an image hypermatrix with a size of m*k based on the k subgraph data sets, wherein a first column in the image hypermatrix corresponds to the first subgraph data set, and m is a number of pieces of subgraph data in the first subgraph data set; 
 randomly generating a weight coefficient for each image in the image hypermatrix; 
 performing row normalization on weight coefficients of images in the image hypermatrix, so that a sum of weight coefficients of each row of images is 1, and a weight coefficient of each image is not greater than W max ; and 
 performing weighted summation on a row of images of the image hypermatrix to obtain a blended image hypermatrix with a size of m*1, wherein images in the blended image hypermatrix are the augmented anonymized image data. 
   
     
     
         18 . The computer system according to  claim 16 , wherein the performing data anonymization processing on the image data based on frequency domain transform includes:
 performing local frequency domain transform processing on the image data to obtain at least one feature map, wherein each feature map of the at least one feature map includes a plurality of elements and corresponds to a data block in the image data, and each element of the plurality of elements corresponds to a frequency in the frequency domain;   constructing, by using elements corresponding to frequencies in the at least one feature map, frequency component channel feature maps corresponding to the frequencies, respectively; and   obtaining at least one target frequency component channel feature map from the frequency component channel feature maps through screening, to obtain the anonymized image data of the image data, wherein a target frequency component channel feature map of the at least one target frequency component feature map is a channel feature for image recognition.   
     
     
         19 . The computer system according to  claim 18 , further comprising:
 after the obtaining the at least one target frequency component channel feature map from the frequency component channel feature maps through screening,
 performing a first shuffle processing on a target frequency component channel feature map to obtain a first shuffled feature map; and 
 performing normalization processing on the first shuffled feature map. 
   
     
     
         20 . A storage medium having executable instructions stored thereon, the executable instructions, when executed by one or more processors, enabling the one or more processors to, individually or collectively, implement acts comprising:
 one or more processors;   one or more storage devices coupled to the one or more processors, the one or more storage devices, individually or collective, having, which when executed by the one or more processors, enable the one or more processors to, individually or collectively, implement acts including:   performing data anonymization processing on image data based on frequency domain transform to obtain anonymized image data of the image data, the anonymized image data including a first subgraph data set, and each subgraph data in the first subgraph data set corresponding to a different frequency; and   performing image blending processing on the subgraph data in the anonymized image data based on data augmentation, to obtain augmented anonymized image data.